AI-guided multi-objective optimization identifies a Jelleine I-derived gram-negative antimicrobial peptide lead.

Antimicrobial peptides are promising alternatives to conventional antibiotics, but their translational development is often limited by the difficulty of simultaneously optimizing antibacterial potency, mammalian-cell compatibility, and formulation-relevant properties. Here, using Jelleine-I (J1) as a representative scaffold, we developed a multi-task learning and reinforcement learning-based framework for the multi-objective optimization of antimicrobial peptides across antibacterial activity, hemolytic toxicity, and self-assembly propensity. Experimental validation of AI-designed analogues identified YB18 (RFRLILRL-NH2) as a balanced Gram-negative-directed lead, with improved antibacterial potency relative to the parent peptide Jelleine-I and MIC values of 8-16 μM against reference Gram-negative strains. YB18 showed low hemolytic activity, acceptable cytocompatibility at antibacterial-relevant concentrations, and membrane-associated bactericidal behavior involving membrane permeability and membrane potential disruption. In an E. coli-infected wound model, topical administration of YB18 reduced the bacterial burden in vivo. Importantly, YB18 retained formulation-relevant self-assembly and formed a viscoelastic hydrogel at 12 mM in 0.8× PBS. The YB18 hydrogel further reduced bacterial counts in infected wounds and showed favorable preliminary local tolerability after repeated topical administration. These findings identify YB18 as a promising peptide lead for local anti-infective applications and support AI-guided multi-objective optimization as an effective strategy for antimicrobial peptide lead discovery.